Interview Xavier Fresquet

Beyond Generative AI: How Interdisciplinary Research Changes Science

Interview with Xavier Fresquet

September 17, 2026

This interview was originally conducted on May 27, 2025.

At the time of the interview, Xavier Fresquet was Deputy Director of the Sorbonne Center for Artificial Intelligence (SCAI). Since February 2026, he has served as Director of SCAI Abu Dhabi at Sorbonne University Abu Dhabi, where he leads research, education, and partnerships in areas including health, environment, machine learning, and the digital humanities. Trained in musicology and digital humanities at Paris-Sorbonne, he joined UPMC, later Sorbonne Université, in 2015. Since 2019, he has helped develop SCAI’s interdisciplinary research and education activities. His work connects AI with the humanities, music, cultural heritage, health, and environmental research

In this interview, Xavier Fresquet explains how PostGenAI@Paris brings AI research into dialogue with physics, law, sociology, ethics, and the humanities. He discusses the people, datasets, educational formats, and institutional structures needed to make AI a meaningful medium of scientific change.

Andreas Sudmann: The PostGenAI@Paris cluster aims to develop ethical, transparent, and socially responsible models. How does this post-generative ambition concretely change the scientific methods and validation practices used in AI research at Sorbonne University, particularly in comparison with the performance-oriented development of generative AI elsewhere?

Xavier Fresquet: Well, I suppose it is a way of articulating our vision and our DNA through what we call a post-generative framework. We are building a research ecosystem in Paris and, more broadly, in France that works more collectively than we did in the past. AI is transforming every domain of science, so we bring together researchers in mathematics and computer science, but also in law, political science, sociology, and ethics. In this way, we are building a more interdisciplinary ecosystem that aims to be transparent and fair. The initiative also involves companies that contribute real-world issues and problems. We do not yet know whether we will achieve everything we aim for, but the objective is to work collectively towards that goal and to do things that smaller teams could not accomplish. The challenge is to organize the ecosystem so that it works. That is the overall objective.

Andreas Sudmann: With initiatives such as physics-informed learning, AI is becoming deeply integrated into specific scientific domains. How does SCAI facilitate an epistemological shift in these fields? In particular, how does it change the ways hypotheses are formulated, evidence is gathered, and discoveries are validated when AI models become central research instruments?

Xavier Fresquet: From my perspective, it is about people. It is about science, of course, but also about people. To bring about an epistemological shift, you need to work with senior researchers who are experts in their fields and open to facilitating change within their communities and their teaching. Part of our role at the university is to teach the next generation how to use these tools, which is a challenge in itself. We facilitate this by giving established researchers the means to attract the right talent—young researchers and postdoctoral researchers who can help them. It is always a question of time and skills. These researchers can help transform fields such as physics, biology, and chemistry. To me, it is fundamentally about people.

Andreas Sudmann: The Collaborative Acceleration Programmes, or CAPs, bring together fields including the humanities and social sciences, as well as law. How do these interdisciplinary structures work in practice? How do they move beyond merely applying AI as a tool and foster a co-evolution of research questions and methods across the technical sciences, humanities, and social sciences?

Xavier Fresquet: That is an important question. We do not aim to create a revolution, because we could not do that. What we can do is help produce shifts on a small, local scale by focusing on a specific problem—for example, a problem in fluid dynamics. The CAPs are somewhat comparable to ERC Synergy Grant frameworks, in which four or five experienced researchers work together on a specific topic. Applying AI in their domains is already significant, because they have to test whether it works. The next question is whether this can foster broader methodological change. AI is then no longer used merely to obtain a result or prediction; it also changes how researchers view their discipline and make new scientific discoveries. This is interesting in physics and, as you discussed with Gérard, in mathematics as well.

Andreas Sudmann: Platforms such as OpenClassrooms and datacraft are presented as important educational partners. Beyond transferring skills, how do these platforms function as epistemic infrastructures? How might they shape the kinds of AI knowledge that are prioritized and the ways students and researchers understand both the possibilities and limitations of AI?

Xavier Fresquet: These two partners are part of an ongoing project that is now approaching its end. We use them for two different reasons. We work with OpenClassrooms because universities do not have enough people to teach even the basics of statistics, programming, AI, or machine learning. If we had enough professors, we would rely entirely on human teachers, but there is an economic and staffing problem. Online courses are therefore a potential solution. They have limitations, of course, but they can meet some of our students’ needs and can also support executive education. The story with datacraft is different. Together, we develop in-person courses for professionals from various sectors, combining interactive sessions with extensive hands-on work. Datacraft’s core activity is to bring together data scientists from companies working in fields such as pharmaceuticals, energy, and transport, give them problems, and enable them to solve those problems collectively. They describe this as a form of digital compagnonnage, referring to the craftspeople who built cathedrals and learned from one another through practice. It is a problem-solving approach, and we incorporate these two educational philosophies into specific courses at the university.

Andreas Sudmann: The AI Lab for the humanities and social sciences aims to understand AI’s societal impact and develop innovative methods. How are its findings and critical perspectives integrated into the core technical development processes at SCAI and PostGenAI@Paris, so that they can influence the design of AI systems from the outset?

Xavier Fresquet: Which lab for the humanities and social sciences do you mean—the one within the AI cluster? That was part of the AI cluster. Perhaps you are referring to the médialab at Sciences Po?

Andreas Sudmann: I am thinking of the médialab at Sciences Po.

Xavier Fresquet: If you mean the médialab at Sciences Po, it is a very interesting and relatively new partner. After our cluster was established, it developed another partnership with OpenAI and began working on democratic commons. The idea is to create well-curated, more balanced, and less biased datasets that are openly available, allowing public and private actors to develop less discriminatory systems. We are integrating the médialab into the cluster and working in that direction. The more we rely on machine-learning systems, the more we need balanced, well-curated, and lawful datasets. They must also respect constraints such as intellectual property. We want to work specifically on these datasets and build less biased AI systems on top of them.

Andreas Sudmann: Given the cluster’s emphasis on digital humanities and large language models, how is it addressing the epistemic challenges that arise when AI is used to analyse large cultural and historical datasets—for example, biases in the data, the opacity of models, and the changing role of human interpretation?

Xavier Fresquet: To me, there are two separate issues. The first is the use of AI in the digital humanities: using it as a tool for humanists working with digital data, whether in music, text, history, or literature. Researchers generally apply existing techniques and models and then compare their outputs critically to determine whether they make sense and can contribute to a scientific investigation. The second issue is the kind of work carried out by the data lab at Sciences Po, or by colleagues in sociology: critically analysing these models as human productions that are necessarily flawed. Perspectives from the social sciences can help improve models that interact with people and therefore need to be balanced, safe, and used with appropriate safeguards. These are two distinct issues.

Andreas Sudmann: Disinformation is another important transdisciplinary topic in AI. From a media-studies perspective, it is also highly relevant to fields such as climatology. How can research and public dialogue around disinformation contribute to an agenda for developing AI technologies that strengthen, rather than undermine, trust and understanding in scientific and public discourse?

Xavier Fresquet: It is an interesting topic. We study disinformation and deepfakes here from several perspectives. From a computer-science perspective, researchers develop systems that assess the quality or authenticity of information, including generated text, video, and sound, and try to determine reliably whether something was produced by a machine rather than a human. Other researchers study disinformation from a network perspective: how false or toxic information spreads, is repeated, and becomes amplified in the digital world. Last week I had a very interesting conversation with a geophysicist from the Institut de Physique du Globe de Paris, just across the street. The institute works with seismic stations around the world, and he had been asked to lead a group on disinformation and the spread of false news on social media following extreme events. He described a false report about a major seismic event in Central Asia, including claims of many deaths, that spread widely on social media even though the event had never occurred. I find it striking that someone trained in geophysics should become involved in work on information and disinformation. Yet this is important because institutions such as his have a responsibility to provide reliable information about what is happening.

Andreas Sudmann: SCAI works with 16 academic and institutional partners and more than 60 industrial partners. Even without directly coordinating all of them, this is a substantial collaborative challenge. What infrastructure—shared data platforms, communication protocols, or collaborative software—is used to manage this network, and how does it shape the flow and character of scientific collaboration?

Xavier Fresquet: To cooperate and collaborate—and sometimes to coordinate—you first need people. Since your last visit, our team has more than doubled in size. For Gérard and me, the work is increasingly managerial: we manage people who themselves manage communication, scientific projects, educational projects, and other activities. This represents a shift in how we work. The university provides digital infrastructure for sharing data and documents and for communicating with our partners. We also use an open-source alternative to Slack for internal communication within specific projects and within the team.

Andreas Sudmann: What is the communication software you use called?

Xavier Fresquet: Mattermost. It is an open-source alternative to Slack and is very convenient. Our needs evolve almost daily or weekly because we are growing quickly, must communicate more, and do not have more hours in the day. We therefore need to become more efficient and continually test new solutions. Even within a team of 20 people, it is challenging to ensure that everyone has the same information.

Andreas Sudmann: You now also have funding opportunities for additional staff. Presumably that helps.

Xavier Fresquet: Yes, but simply coordinating the 20 people here is difficult, especially when trying to give everyone the same level of information. We communicate as much as we can and are not bad at it, but we constantly need strategies to keep everyone on the same page, circulate information quickly, and prevent anything from being lost. People should not feel excluded. You have to find time to communicate with them; that is part of a management role. Otherwise, people become unhappy and leave, and we want them to stay.

Andreas Sudmann: This brings me back to our discussion over lunch about how much AI research is project-driven rather than supported by permanent structures. SCAI is itself a structure that develops projects, but would it also be important to maintain continuing research that is not limited to the typical three- or four-year duration of a project?

Xavier Fresquet: That already happens, but SCAI is a different kind of structure. Long-term research is developed within established laboratories—the robotics and mathematics laboratories, for example—which have existed for decades and define their own research strategies. Our role is to coordinate specific project-driven initiatives. We bring together the appropriate researchers to work on particular large-scale projects. We do not conduct basic research internally at SCAI. Instead, we help research develop within the laboratories by providing resources that support their own strategies.

Andreas Sudmann: PostGenAI@Paris aims to define, or at least guide the definition of, post-generative AI. What role do media and communication strategies play in establishing this concept within the diverse cluster and across the international scientific community? How might they help turn it into a recognized and influential paradigm?

Xavier Fresquet: That is an interesting point. We coined the term before we had fully defined it. We believed there was an idea or perhaps a paradigm behind it, and expected it to take shape over time. I think that is now happening, and media and communication studies clearly have a role to play. At the launch of PostGenAI@Paris in April, we invited Daniel Andler, a philosopher and professor emeritus at Sorbonne University, to reflect on what post-generative AI might mean. He has witnessed the development of AI from the era of Jacques Pitrat and symbolic AI through deep learning and generative models. His argument was that the technology is already shifting again. Even within generative models and agents, rules are returning: physics-informed machine learning incorporates physical laws, while agentic systems again make use of symbolic rules. Five years after the rise of today’s generative systems, we are changing how these systems are produced and used, and another shift will probably follow. In that sense, we are already entering a post-generative paradigm. We are no longer simply using GANs or diffusion models to generate data; we have moved a step further, and these systems are already being implemented in companies and everyday life. Most people may not recognize that transition because the public discussion still centres on generative AI.

Specialists in information science and media studies should therefore investigate this shift critically and help citizens understand that we have entered a new phase.

Andreas Sudmann: This is an interesting question in itself. One could argue that hybrid systems combining deep learning with symbolic reasoning have existed throughout the recent AI boom. I was therefore wondering specifically about the “post” in post-generative AI. The term “post-digital,” for example, can describe a condition in which digital technology has become ubiquitous and ordinary. That does not seem fully applicable to AI yet, even though AI has entered many everyday computer-based routines. Does “post-generative” carry a similar implication, or does it instead describe a return to and recombination with earlier approaches?

Xavier Fresquet: Yes and no. We are still moving forward, but the techniques are changing, evolving, and becoming increasingly entangled. Your comparison is useful if “post-generative AI” describes a world in which generative AI is no longer perceived as something new but has become part of everyday life. I think we are entering that phase because generative systems are being used more and more. They are changing how we interact with digital systems and search engines. Our generation learned to search by selecting keywords—not because that is the natural way to formulate a question, but because search engines required it. For us, searching often means finding the three words most likely to lead to the right website. Our children will interact differently. They will ask natural-language questions and increasingly use systems such as Perplexity. They will not need to identify the exact keywords that match a website. We are not fully there yet, but within a few years the way younger people search for information online will be completely different from ours.

Andreas Sudmann: It is interesting that you refer to the keywords we learned to use in search engines, rather than to the earlier practice of going to a library and searching differently.

Xavier Fresquet: Of course, one can still go to a library. But search engines arrived and, because we wanted to find information through them, they forced us to think in a way that is rather unnatural for human beings. The next generation will find it much more natural to interrogate the web through ordinary language. When you look at the kinds of queries we have learned to type into computers, they are actually quite strange.

Andreas Sudmann: It is also interesting to consider the still-emerging practice of prompt engineering, which some people even describe as an art form. Prompts are remarkably flexible: even weak prompts can sometimes produce good results, unlike the keywords used in traditional search. Yet effective prompting still requires refinement, precision, repetition, and variation; it is a serial practice. Do you expect natural-language interaction with computers to replace classical coding, as many predict? And do you sense anxiety that established capabilities—coding, domain expertise, and rigorous research practices—may erode when a relatively naïve general question can already produce a plausible result? Let us begin with the first part: prompting versus coding.

Xavier Fresquet: Yes—the coding question is very interesting. We are moving towards a situation in which almost no one writes everything from scratch. People adapt and extend existing code, and we increasingly use generative tools to teach students how to code. Gérard recently told me about a lecture by Yuval Noah Harari in Paris in which the possibility was raised that future systems may make coding unnecessary because people will interact with computers entirely through natural language. I do not yet see that future clearly, but many people do. Some coding expertise will probably remain necessary, although I am not certain. The disappearance of coding would imply that we had achieved something like a perfect computing language, which is itself questionable because languages change so quickly. Daniel Andler approached the issue from another direction when discussing the skills required by the next generation. He proposed a thought experiment: imagine a complete blackout in which electricity and digital systems no longer work. What skills would people need to survive and rebuild society? Those are the foundational skills we should teach the next generation. Students also need to learn how to use generative systems, but universities may not need to assess every one of those operational skills. What we should assess are the deeper capacities that remain necessary even in a non-digital world.

Andreas Sudmann: It is striking that you frame this as a thought experiment about the skills required in a time of crisis.

Xavier Fresquet: Yes, it is a scenario of crisis, although we are not claiming that it will happen. The point is to distinguish among the skills students need, the skills universities can meaningfully assess, and those that remain important even if they are not directly assessed by an academic institution.

Andreas Sudmann: As a final question, and in view of the global competition for talent and the ambition of European technological sovereignty, how does the Sorbonne cluster use its distinctive focus on ethical AI integrated with the humanities and social sciences as a strategic advantage? How is that focus communicated in order to attract and retain researchers motivated by these scientific and societal goals?

Xavier Fresquet: Funding is distributed unevenly across disciplines. There is less funding in musicology than in computer science, and it is generally more difficult to finance work in sociology, law, and the humanities and social sciences. I believe we received this grant from the French government partly because we deliberately included projects from those fields. The grant now enables us to fund projects that can attract researchers for two, three, four, or five years. We hope that at least half of them will remain and continue building the university’s capacity for the coming era. That is how the cluster can help.

Andreas Sudmann: One final question: after submitting the current application, did anything emerge that made you think it should have been included? Can you give an example?

Xavier Fresquet: One issue resulted from the structure of the application: we had to define projects in which public and private partners worked together. We succeeded in doing that for 21 large projects, but could not include everything. Ocean science, for example, is Sorbonne University’s second-largest discipline after mathematics, and we have no project devoted to it, even though the university is a European leader in the field and has outstanding researchers and data. Because the application timetable was so short, we could not identify one or two companies with which our marine stations could develop a joint project. That is a real regret, although we will develop another project with them. I also regret that we did not propose a stronger project on AI for mathematics. We have a mathematics project, but it focuses more on the theory of machine learning. A project using AI to solve mathematical problems, along the lines of what you discussed with Gérard, would have been important. The same is true of theoretical approaches to quantum machine learning, an area that is already emerging. Finally, we could have included many more health topics. The projects we do have are excellent, but with more time we could have incorporated much more.

Andreas Sudmann: The broader lesson may be that a portion of funding should remain flexible, so that its precise purpose can be determined later.

Xavier Fresquet: We do have some flexible funding, but not at the scale of these projects, which receive two or three million euros each for a dedicated programme and therefore have much greater impact. Still, we will support the omitted areas and pursue other opportunities. There are many possibilities at the moment; for example, we are already preparing a major medical project for next September. We continue to build new projects even after this large application, so I am not worried overall. But I still regret the absence of an ocean-science project.

Andreas Sudmann: Thank you very much.